ARIA: An Autonomous Reasoning and Intervention Architecture for Bioregenerative Life Support Systems on Long-Duration Mars Missions

Long-duration crewed Mars missions impose a fundamental constraint that is absent from all prior human spaceflight: Earth-to-Mars one-way communication delays of 4 to 24 minutes render real-time mission control support impossible during life support system (LSS) emergencies. Simultaneously, the next generation of Mars habitat life support integrates bioregenerative subsystems, including closed biological loops involving microalgae, higher plants, and bacterial compartments, whose emergent, non-linear behaviour exceeds the diagnostic capacity of classical deterministic fault detection and isolation (FDI) methods. We present \ARIA{} (Autonomous Reasoning and Intervention Architecture), a five-layer hybrid framework that augments classical ECLSS fault detection with a network of specialised reasoning agents and a retrieval-augmented generation (RAG) large language model (LLM) orchestrator. \ARIA{} provides ranked natural-language intervention options to the crew while enforcing a hard human-confirmation boundary: no LLM output actuates directly on hardware. We formalise the system state space, per-subsystem anomaly severity metrics, inter-agent cascade risk quantification, and four safety constraints governing system behaviour. A detailed hypothetical case study, involving concurrent Carbon Dioxide Removal Assembly (CDRA) degradation and algae bioreactor pH collapse during a communication blackout with 19-minute one-way delay, demonstrates that \ARIA{} identifies a coupled cross-subsystem intervention that is unlikely to be found by a fatigued crew under unaided time pressure. Under the \ARIA{}-assisted trajectory, cabin CO$_2$ partial pressure peaks at 6.4~mmHg (below the 7.6~mmHg cognitive impairment threshold) and bioreactor culture is preserved. Under the unassisted counterfactual, the bioreactor culture is projected to crash irreversibly before Earth contact is re-established. \ARIA{} represents, to the authors' knowledge, the first formal agentic LLM framework designed specifically for bioregenerative LSS management, and contributes a safety boundary taxonomy applicable to AI-assisted decision support in other safety-critical domains.